Reproduce “FSDR: Frequency Space Domain Randomization for Domain Generalization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiaxing Huang, Dayan Guan, Aoran Xiao, Shijian Lu
Nanyang Technological University researchers introduced Frequency Space Domain Randomization (FSDR), a method that processes images in the frequency domain to separate domain-invariant and domain-variant features for selective randomization, achieving an average mIoU of 43.1 on GTA5-to-real-world semantic segmentation tasks, surpassing state-of-the-art domain generalization methods by over 3 percentage points.
Key insights
- Images contain both domain-variant (e.g., textures, colors) and domain-invariant (e.g., object shapes) features, and selectively randomizing only domain-variant features improves generalization.
- Operating in the frequency domain allows for more granular control over image properties, enabling explicit separation and differential treatment of domain-invariant and domain-variant components.
- Dynamically learning which frequency components are domain-invariant or domain-variant through an iterative process (FSDR-SL) leads to more effective randomization than static empirical identification.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.